l0llm Pure C# & Intel Level Zero LLM Runner & REST Server
Ultra-fast, zero-allocation native LLM runner and OpenAI-compatible REST server engineered exclusively for Intel® GPUs. Built entirely in managed C# on top of LevelZero.NET with zero Python runtime, zero Conda environments, and zero heavyweight native frameworks.
Intel Iris® Xe Mobile Performance (15W TDP)
Measured on Intel® Iris® Xe Graphics (96 EU, 15W TDP mobile) running Qwen 2.5 with direct Level Zero command lists.
Fast sustained generation on lightweight mobile integrated GPUs.
Immediate first-token responsiveness with minimal queue lag.
Zero Garbage Collection pressure during continuous token generation loops.
Iris Xe, Arc A-Series, Lunar Lake, Arrow Lake, Battlemage, Flex/Max.
Zero-Friction Local & Server-Side LLM Deployment
Pure Managed C# Pipeline
Direct C# bindings directly to Level Zero driver exports. No PyTorch, no ONNX Runtime C++ dependencies, and no bloated runtime installs.
OpenAI-Compatible REST API
Drop-in local server supporting /v1/chat/completions with Server-Sent Events (SSE) streaming, /v1/models, and built-in CORS for Open-WebUI, Chatbox, and LangChain.
Native SafeTensors Support
Reads standard Hugging Face model.safetensors and sharded indexes (model.safetensors.index.json) with zero-copy memory mapping.
Wide Model Family Support
Out-of-the-box optimized execution for LLaMA 3/3.1/3.2, Qwen 2.5/3, DeepSeek, Mistral, and Phi-3.5 architectures.
Interactive CLI & REST Server
# Start high-performance OpenAI server on Intel GPU
l0llm serve --model ./models/qwen2.5-7b --port 8080 --host 0.0.0.0
# Interactive single-turn or streaming prompt
l0llm run --model ./models/llama3.2-3b --prompt "Explain quantum computing in three sentences"Downloads & Verification
| Asset | Target Platform | Version | Action |
|---|---|---|---|
| l0llm-win-x64-v2.0.0.zip | Windows x64 (Intel GPUs) | v2.0.0 | Download |